Skip to main content
SupaCortex
Sign in Book a Demo
Feature

The best answers your team has written are already in the ticket history

Agents have explained the same fix a hundred times. SupaCortex reads resolved conversations, proposes the question and the answer behind each one, and adds it to your knowledge only after a person approves it.

Book a demo See pricing
  • Zendesk, osTicket, UnifyDesk
  • Redacted before analysis
  • Nothing searchable until approved

How does SupaCortex learn from past tickets?

A ticket feed reads resolved conversations from Zendesk, osTicket or UnifyDesk. Each conversation is checked for consent and against a category deny-list, has personal data replaced with placeholders, and is analysed by the model to propose a question-and-answer pair. Rule checks and de-duplication against existing knowledge remove the weak and the repeated. What is left lands in a review queue, and only a candidate a person approves becomes a document that can be searched and cited.

  • Sources: Zendesk, osTicket (read-only database) and UnifyDesk.
  • Emails, phones, IDs, amounts and names redacted before the model reads anything.
  • Duplicates of existing knowledge and of earlier candidates are dropped.
  • Side-by-side review: the redacted excerpt next to the editable Q&A.
  • Approve, edit or reject one at a time, or in bulk.
  • Approved answers become normal, citable documents; delete one and it is withdrawn.
How it works

From closed ticket to approved answer

  1. 01

    Filter

    Only resolved conversations with a human reply, outside deny-listed categories, from a source that has opted in.

  2. 02

    Redact

    Names, contact details, account and ID numbers and amounts become placeholders. Error codes and setting names stay.

  3. 03

    Propose

    The model drafts the question and the answer; rule checks and de-duplication discard what is weak or already known.

  4. 04

    Approve

    A reviewer reads the excerpt beside the proposal, edits if needed, and approves. Only then is it indexed.

In detail

Careful by default, because these are customers' conversations

Privacy before analysis, not after

The model never reads a raw conversation. Redaction happens first, and the rules check its work afterwards.

  • Emails, phone numbers (Indian and international), card numbers and IP addresses.
  • PAN, Aadhaar, UAN, IFSC, bank account digits and employee IDs.
  • Amounts and pay figures, and every name the source or the text gives away - in every message and the subject.
  • A candidate with personal data still in it fails the checks and never reaches review.

A review queue built for volume

A desk with years of history produces a lot of candidates. Reviewing them should not take years too.

  • Ordered by the model's confidence, so the clearest candidates come first.
  • Filter by status, feed or text; keyboard shortcuts to approve, reject, edit and move on.
  • Bulk approve or reject a selection, or everything matching a filter.
  • A summary of pending, approved and rejected counts, the approval rate and recent runs.

A conversation mined again replaces its earlier pending candidate. It never touches one you have already reviewed.

Kept in its place

A fix that worked for one customer last year is useful evidence - but it should not outrank the manual.

  • Customer-specific answers can stay scoped to that customer.
  • Official documents outrank mined answers in search.
  • Deleting an approved answer's document withdraws it from answers straight away; if it is mined again, it goes back to review.
  • Only Editors and above can review; Viewers can see the queue.

Related

Turn Tickets Into a Knowledge Base

The answers your agents have already written, reviewed once and reused for every customer after.

Reply Drafts for Help Desks & Email

Grounded, cited reply drafts inside UnifyDesk or any desk on the Draft API. A person still sends.

Answers With Citations

Hybrid search, a confidence gate and checked citations - it says "I don't know" instead of guessing.

AI Replies Inside UnifyDesk

Agents get a cited draft for every ticket - drafted when the customer writes, sent only by a person.

Document Uploads

Drop in PDFs, Word and Excel files, Markdown and HTML in bulk - or a zip of them. Each is parsed and indexed.

Security

Separate databases per workspace, encrypted secrets, and answers scoped to who is asking.
Frequently asked

Questions people ask before they start

Zendesk, osTicket and UnifyDesk. Zendesk connects with an API token, osTicket with a read-only database connection, and UnifyDesk through its own export, which the desk has to switch on first.

No. Every mined question and answer waits in a review queue. Nothing is searchable until a person on your team approves it, and they can edit the wording first or reject it with a note.

Before the model sees a conversation, emails, phone numbers, card numbers, IP addresses, PAN, Aadhaar, IFSC and account numbers, employee IDs, amounts and people's names are replaced with placeholders. A candidate that still contains personal data after that fails the rule checks and is not offered for review.

Open conversations, conversations with no reply from a human agent, anything in a category on the deny-list (plus your own list), and - for UnifyDesk - any desk that has not opted in. Internal notes never reach the analysis.

Below them. An approved answer is indexed as a standard document, and anything you mark as community or past-ticket content is ranked lower still and dropped altogether when enough official evidence exists.

Business and Enterprise. Mining one conversation uses 2 AI credits.

Find out what your ticket history already knows.

Book a 30-minute demo on your own documents. We set up a 14-day trial workspace for you - no credit card, no setup fee.

Book a demo See pricing

Questions first? Email info@greatdevelopers.com or call +91 8585953085.

Still curious? Ask your favorite AI about SupaCortex.

ChatGPT Claude Perplexity Grok